TLDR: A new study explores how large language models (LLMs) integrated into social robots handle user privacy. It found that while LLMs generally favor privacy, their understanding often lacks human nuance. However, providing LLMs with examples of human preferences (few-shot prompting) significantly improved their alignment with user expectations. The research suggests that future social robots need to learn individual privacy preferences through interaction to become truly privacy-aware controllers.
As social robots become more integrated into our daily lives, especially within private spaces like homes, the question of how they handle personal data becomes increasingly critical. These embodied agents, designed to interact with people using verbal and non-verbal cues, are evolving rapidly with the integration of large language models (LLMs). While LLMs offer new opportunities for enhanced human-robot interaction, they also necessitate the collection of sensitive data, including audio, images, video, and location information. This creates a significant tension between the utility these robots provide and the potential privacy risks they pose.
A recent study, titled “Benchmarking LLM Privacy Recognition for Social Robot Decision Making,” delves into this crucial area. Conducted by researchers Dakota Sullivan, Shirley Zhang, Jennica Li, Heather Kirkorian, Bilge Mutlu, and Kassem Fawaz from the University of Wisconsin – Madison, the paper explores the extent to which off-the-shelf LLMs are privacy-aware in the context of household social robots. The research highlights that while LLMs can enhance robot capabilities, their ability to manage sensitive data appropriately is not yet fully understood.
The researchers developed a set of privacy-relevant scenarios, grounded in the concept of Contextual Integrity (CI), to evaluate both human and LLM privacy preferences. They first surveyed 450 users to understand their privacy expectations regarding in-home social robot behaviors and how their personal privacy orientation influenced these choices. Subsequently, the same scenarios and questions were presented to ten state-of-the-art LLMs.
The findings revealed a notable disconnect: the agreement between human privacy preferences and LLM responses was surprisingly low. This suggests that while LLMs might have a broad understanding of privacy sensitivity, they may lack the nuanced awareness required for complex, real-world social robot interactions. To investigate how LLMs could potentially act as privacy controllers, the study implemented four additional prompting strategies, including providing LLMs with user privacy orientation scores, scenario sensitivity ratings, and a combination of both. The most effective strategy, few-shot prompting (where LLMs were given an example of a participant’s responses to a scenario), significantly improved the models’ alignment with human preferences.
This research underscores that humans are generally highly privacy-oriented and consistently prefer privacy-preserving robot behaviors, even in situations deemed minimally sensitive. LLMs, by default, tend to lean towards privacy-preserving responses, but their reasoning for these choices often lacks the depth of human understanding. The study also observed that larger, closed-source models generally performed better than their open-source counterparts, though even the best-performing models could ‘overcorrect’ when given too much specific information.
The implications of this work are significant for the future design of LLM-powered social robots. It suggests that for robots to truly respect user privacy, they need to be able to learn and adapt to individual user preferences over time through natural interactions. This could involve a calibration phase or continuous learning, allowing the robot to understand the unique nuances of a household’s privacy expectations. Furthermore, the study raises concerns about the use of non-local (cloud-based) LLMs due to data privacy implications, proposing solutions like clear user data policies or the development of sufficiently privacy-aware local models.
Also Read:
- AI’s Achilles’ Heel: Why More Feedback Can Harm Large Language Model Performance
- Unmasking Hidden Biases: The Subtle Ways Language Models Change Facts
While the study provides valuable insights, it also acknowledges limitations, such as potential biases in participant privacy orientation scores compared to prior research, and the ongoing challenge of definitively concluding whether LLMs are truly privacy-aware. Nevertheless, this work paves the way for developing social robots that can act as more rational and user-aligned privacy controllers, bridging the gap between technological utility and human privacy expectations. You can read the full research paper here: Benchmarking LLM Privacy Recognition for Social Robot Decision Making.
